ABSTRACT
Addressing missing data often involves the imputation of missing values. The statistical trick with Bayesian imputation is to model the variable that has missing values. Each missing value is assigned a unique parameter. The observed values give us information about the distribution of the values. The right-hand plot shows the inferred relationship between the two predictors. The procedure performs multiple draws from an approximate posterior distribution of the missing values, performs separate analyses with these draws, and then combines the analyses in a way that approximates full Bayesian imputation. Multiple imputation was developed in the context of survey non-response, and it actually has a Bayesian justification.
